ICLR 2026 Papers — Page 10
International Conference on Learning Representations · 5356 papers
Controlling Repetition in Protein Language Models
Jiahao Zhang (Mohamed bin Zayed University of Artificial Intelligence), Lijie Hu (King Abdullah University of Science and Technology)
GenerationData SynthesisProtein Structure PredictionTransformerContrastive LearningBiomedical Data
🎯 What it does: This paper systematically studies the pathological repetition problem occurring during the generation process of protein language models (PLM), and proposes a set of quantitative metrics to evaluate repetition degree and structural feasibility; subsequently, it introduces the 'Utility-Controlled Contrastive Steering (UCCS)' method based on contrastive learning, which reduces repetition during inference by injecting contrast vectors while maintaining or enhancing AlphaFold structural confidence.
Converge Faster, Talk Less: Hessian-Informed Federated Zeroth-Order Optimization
Zhe Li (Rochester Institute of Technology), Haibo Yang (Rochester Institute of Technology)
OptimizationFederated LearningLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes HiSo, a zeroth-order federated optimization algorithm that utilizes a global diagonal Hessian approximation, significantly accelerating LLM federated fine-tuning while ensuring scalar-only communication.
Convergence Analysis of Tsetlin Machines under Noise-Free and Noisy Training Conditions: From $2$ Bits to $k$ Bits
Xuan Zhang (University of Agder), Ole-Christoffer Granmo (University of Agder)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper provides a systematic theoretical analysis of the convergence of the Tsetlin Machine (TM). It first presents convergence proofs for noise-free and noise-perturbed training scenarios of 2-bit AND, OR, and XOR logic operations, and further generalizes these results to arbitrary k-bit cases. Additionally, it investigates the impact of uncertain labels and irrelevant variables on TM convergence, and provides corresponding convergence conditions.
Convergence Dynamics of Over-Parameterized Score Matching for a Single Gaussian
Yiran Zhang (Tsinghua University), Simon Shaolei Du (University of Washington)
OptimizationScore-based Model
🎯 What it does: Studied the convergence dynamics of gradient descent learning a single Gaussian distribution under an over-parameterized score-matching framework
Convergence of an actor-critic gradient flow for entropy regularised MDPs in general spaces
Denis Zorba (University of Edinburgh), Lukasz Szpruch (University of Edinburgh)
Reinforcement LearningStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies the stability and global convergence of actor-critic gradient flows in Markov decision processes (MDPs) with general state and action spaces under entropy regularization, through continuous-time limit analysis;
Convergence of Muon with Newton-Schulz
Gyu Yeol Kim (Seoul National University), Min-hwan Oh (Seoul National University)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper provides a theoretical analysis of the optimizer MUON, which uses matrix-structured parameters in deep learning, and proves that its finite-step Newton-Schulz approximation orthogonalization version converges to steady-state points in non-convex optimization. It demonstrates that the convergence rate matches the ideal SVD-polar variant and quantifies the double-exponential decay of Newton-Schulz error with step number q. Experiments on multiple models and datasets validate the theory.
Convergence of Regret Matching in Potential Games and Constrained Optimization
Ioannis Anagnostides (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
Optimization
🎯 What it does: The paper studies the convergence properties of Regret Matching (RM) and its improved version RM+ in potential games and constrained optimization problems, proving that RM+ can act as a first-order optimizer to rapidly converge to ε-KKT points, while providing a lower bound for RM's exponential slow convergence in certain potential games.
Convergent Differential Privacy Analysis for General Federated Learning
Yan Sun (University of Sydney), Dacheng Tao (Nanyang Technological University)
Federated LearningSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: Studied the convergence of differential privacy in federated learning, provided the worst-case f-DP privacy analysis for Noisy-FedAvg and Noisy-FedProx, and proved privacy convergence under non-convex smooth objectives;
Convex Dominance in Deep Learning I: A Scaling Law of Loss and Learning Rate
Zhiqi Bu (Meta Superintelligence Labs), Jialin Mao (Independent Researcher)
OptimizationImageTextMultimodality
🎯 What it does: Studies the convexity behavior in deep learning, derives and verifies the scaling laws between learning rate and loss, and proposes a two-dimensional scaling model capable of predicting optimal learning rates and loss across different training durations and model scales.
Convex Efficient Coding
Will Dorrell (University College London), James C. R. Whittington (University of Oxford)
OptimizationRepresentation Learning
🎯 What it does: This paper proposes converting the neural coding problem into convex optimization on a representation similarity matrix, thereby obtaining an analytical optimal coding solution and deriving a series of theoretical conclusions.
ConvT3: Structured State Kernels for Convolutional State Space Models
Jaeyoung Hong (SolverX), Minseon Gwak (University of Massachusetts Amherst)
GenerationConvolutional Neural NetworkVideoPhysics Related
🎯 What it does: This paper proposes the ConvT3 model, which expands the state kernel of ConvSSM from 1×1 to 3×3 to more comprehensively capture spatiotemporal dynamics.
Cooperative Sheaf Neural Networks
André Ribeiro (Getulio Vargas Foundation), Diego Mesquita (Getulio Vargas Foundation, 2 δ AI)
ClassificationComputational EfficiencyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Propose Cooperative Sheaf Neural Networks (CSNN), a neural network based on a cellular layered structure with directed graph orientation, enabling nodes to adaptively decide on cooperative behaviors for information transmission and reception;
COOPERTRIM: Adaptive Data Selection for Uncertainty-Aware Cooperative Perception
Shilpa Mukhopadhyay (University of California, Riverside), Hang Qiu (University of California, Riverside)
Autonomous DrivingComputational EfficiencyTransformerReinforcement LearningPoint CloudBenchmark
🎯 What it does: Propose an adaptive feature selection framework named COOPERTRIM, designed for multi-vehicle collaborative perception to transmit only the most important and necessary features for the target task, thereby significantly reducing communication bandwidth consumption.
CoPRS: Learning Positional Prior from Chain-of-Thought for Reasoning Segmentation
Zhenyu Lu (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Yaowei Wang (Meituan)
SegmentationTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose an end-to-end multimodal chained reasoning segmentation framework, CoPRS, which converts language reasoning results into dense differentiable heatmap location priors through learnable condensed queries, and generates fine-grained segmentation masks using a lightweight decoder.
Copy-Paste to Mitigate Large Language Model Hallucinations
Yongchao Long (Tianjin University of Technology), Shenda Hong (Peking University)
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a 'Copy-Paste' generation paradigm, directly copying fragments from retrieved contexts to generate answers, to enhance contextual faithfulness in Retrieval-Augmented Generation (RAG) systems; and construct a two-stage training framework: Stage-1 generates high-repetition-rate candidate answers through three prompts (CP-Order, CP-Link, CP-Refine); Stage-2 performs Direct Preference Optimization (DPO) using automatically generated high-repetition-rate preference dialogues, resulting in CopyPasteLLM.
CoRA: Boosting Time Series Foundation Models for Multivariate Forecasting through Correlation-aware Adapter
Hanyin Cheng (East China Normal University), Chenjuan Guo (East China Normal University)
Computational EfficiencyTransformerSupervised Fine-TuningContrastive LearningTime Series
🎯 What it does: Propose a lightweight CoRA plugin that significantly enhances multivariate time series forecasting performance during the fine-tuning stage by leveraging the internal representations and prediction results of TSFM.
CORDS - Continuous Representations of Discrete Structures
Tin Hadži Veljković (UvA-Bosch Delta Lab University of Amsterdam), Jan-Willem van de Meent (UvA-Bosch Delta Lab University of Amsterdam)
Object DetectionGenerationRepresentation LearningDrug DiscoveryConvolutional Neural NetworkTransformerFlow-based ModelImagePoint CloudGraphTime Series
🎯 What it does: This paper proposes CORDS (Continuous Representations of Discrete Structures), a reversible continuous-domain representation that encodes arbitrary-sized discrete sets into density fields and feature fields, learning directly in the field domain and precisely decoding back to discrete sets.
CORE: Concept-Oriented Reinforcement for Bridging the Definition–Application Gap in Mathematical Reasoning
Zijun Gao (University of Illinois Urbana Champaign), Ben Zhou (Arizona State University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the CORE framework, enhancing the concept recognition and application capabilities of LLMs in mathematical reasoning through concept-oriented reinforcement learning.
Corner Gradient Descent
Dmitry Yarotsky (Applied AI Institute)
OptimizationImage
🎯 What it does: Proposed a new infinite-memory SGD algorithm based on complex-plane contours called Corner Gradient Descent (Corner SGD), and proved that under the power-law spectrum conditions of infinite-dimensional quadratic problems, the convergence exponent can be accelerated from O(t^{-ζ}) to O(t^{-θζ}), where θ can approach 2, achieving optimal acceleration.
Correlated Policy Optimization in Multi-Agent Subteams
Dingyang Chen (Amazon), Qi Zhang (Worcester Polytechnic Institute)
OptimizationReinforcement LearningBenchmark
🎯 What it does: Propose utilizing Bayesian network subteam structures to achieve local relevance strategies in multi-agent collaborative learning, along with theoretical convergence and experimental validation.
Cortical Policy: A Dual-Stream View Transformer for Robotic Manipulation
Xuening Zhang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
Robotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelImageVideo
🎯 What it does: Proposes Cortical Policy—a dual-stream view transformer that enhances 3D spatial reasoning using a static perspective, achieves adaptive trajectory planning via a dynamic perspective (based on a human gaze estimation model), and fuses the two features for robotic manipulation.
CortiLife: A Unified Framework for Cortical Representation Learning across the Lifespan
Pengcheng Xue (Nanjing University of Aeronautics and Astronautics), Xuyun Wen (Nanjing University of Aeronautics and Astronautics)
ClassificationRepresentation LearningTransformerPrompt EngineeringVision Language ModelAuto EncoderContrastive LearningBiomedical DataComputed TomographyAlzheimer's Disease
🎯 What it does: Proposes a unified CortiLife framework for representation learning of the human cortical surface across the entire lifespan.
COSA: Context-aware Output-Space Adapter for Test-Time Adaptation in Time Series Forecasting
Jeonghwan Im (Seoul National University of Science and Technology), Hyuk-Yoon Kwon (Seoul National University of Science and Technology)
Domain AdaptationTime SeriesBenchmark
🎯 What it does: Proposed a test-time adaptation module called COSA, which directly performs residual correction on the outputs of a frozen baseline model in time series prediction, thereby improving prediction accuracy in non-stationary environments.
COSMO-INR: Complex Sinusoidal Modulation for Implicit Neural Representations
Pandula Thennakoon (University of Peradeniya), Vijitha R. Herath (University of Peradeniya)
RestorationSuper ResolutionRepresentation LearningConvolutional Neural NetworkNeural Radiance FieldImagePoint Cloud
🎯 What it does: Proposed a new INR activation function called COSMO-RC, combining complex sinusoidal modulation with Raised Cosine activation to achieve more complete spectral support, enhancing INR performance across multi-task scenarios.
Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning
Moo Jin Kim (NVIDIA), Jinwei Gu (NVIDIA)
Robotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelDiffusion modelWorld ModelVideo
🎯 What it does: Fine-tune the pre-trained video generation model Cosmos-Predict2 into a unified robot control policy that can directly generate actions, predict future states and values, and support planning-based decision-making;
COSMOS: A Hybrid Adaptive Optimizer for Efficient Training of Large Language Models
Liming Liu (Georgia Tech), Tuo Zhao (Georgia Tech)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a new hybrid adaptive optimizer called COSMOS for efficiently training large language models.
Cost-of-Pass: An Economic Framework for Evaluating Language Models
Mehmet Hamza Erol (Stanford University), James Zou (Stanford University)
Computational EfficiencyTransformerText
🎯 What it does: This paper introduces the 'cost-of-pass' metric, which combines the accuracy of language models with inference costs to measure the economic efficiency of models. Based on this metric, it defines frontier cost, tracks technological progress, analyzes the contributions of model families, and evaluates the economic value of inference technologies.
CoT Vectors: Transferring and Probing the Reasoning Mechanisms of LLMs
Li Li (Southeast University), Xu Yang (Southeast University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes extending the concept of task vectors to multi-step reasoning, designing CoT vectors and demonstrating how to extract or learn these vectors from the support set to guide reasoning during LLM forward propagation.
CoT-Evo: Evolutionary Distillation of Chain-of-Thought for Scientific Reasoning
Kehua Feng (Zhejiang University), Huajun Chen (Zhejiang University)
Knowledge DistillationTextBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a chain-of-thought (CoT) distillation framework called COT-EVO based on evolutionary algorithms, aiming to generate high-quality, domain-specific scientific reasoning paths and train smaller models.
CoT-RVS: Zero-Shot Chain-of-Thought Reasoning Segmentation for Videos
Shiu-hong Kao (Hong Kong University of Science and Technology), Chi-Keung Tang (Hong Kong University of Science and Technology)
Object TrackingSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextChain-of-Thought
🎯 What it does: Propose a zero-training, unsupervised CoT-RVS framework that leverages the chain-of-thought (CoT) capabilities of multimodal large language models to perform spatiotemporal reasoning on videos and generate instance-level mask sequences;
CounselBench: A Large-Scale Expert Evaluation and Adversarial Benchmarking of Large Language Models in Mental Health Question Answering
Yahan Li, Ruishan Liu (University Of Southern California)
Adversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed a large-scale mental health Q&A benchmark COUNSELBENCH, integrating expert-evaluated LLM responses and adversarial questionnaires;
Count Bridges enable Modeling and Deconvolving Transcriptomic Data
Nic Fishman (Harvard University), Omar Abudayyeh (Brigham and Women's Hospital)
Data-Centric LearningDiffusion modelScore-based ModelBiomedical Data
🎯 What it does: Propose an integer bridge model based on the Poisson birth-death process for generating and deconvoluting count data.
Count Counts: Motivating Exploration in LLM Reasoning with Count-based Intrinsic Rewards
Xuan Zhang (Fudan University), Yuan Qi (Fudan University)
Large Language ModelReinforcement LearningText
🎯 What it does: Propose the MERCI algorithm, which promotes exploration and diversity in multi-step reasoning by designing a counting-based intrinsic reward mechanism for the LLM inference process;
Counterfactual Explanations on Robust Perceptual Geodesics
Eslam Zaher (ARC Training Centre for Information Resilience), Fred Roosta (ARC Training Centre for Information Resilience)
OptimizationExplainability and InterpretabilityAdversarial AttackGenerative Adversarial NetworkImageAgriculture Related
🎯 What it does: Propose a latent space geometric path optimization method called PCG based on a robust perceptual Riemannian metric to generate semantically coherent adversarial explanations.
Counterfactual LLM-based Framework for Measuring Rhetorical Style
Jingyi Qiu (University of Michigan), Zongyi Li (MIT)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Generate paper abstracts with identical content but varying rhetoric using LLM role generation, quantify the rhetorical strength of ML papers through pairwise comparisons judged by LLMs and calibrated with the Bradley-Terry model.
Counterfactual Reasoning for Retrieval-Augmented Generation
Huaiyu Qin (Renmin University of China), Yunhai Wang (Renmin University of China)
GenerationRetrievalExplainability and InterpretabilityTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the CF-RAG framework, achieving causal reasoning in retrieval-augmented generation through counterfactual exploration and parallel arbitration, thereby escaping the relevance trap.
Counterfactual Structural Causal Bandits
Min Woo Park (Seoul National University), Sanghack Lee (Seoul National University)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper introduces achievable counterfactual actions into the structural causal bandit framework, defines CTF-MIS and CTF-POMIS, and proposes an efficient algorithm for enumerating representative CTF-POMIS;
Coupled Transformer Autoencoder for Disentangling Multi-Region Neural Latent Dynamics
Ram Dyuthi Sristi (UC San Diego), Gal Mishne (UC San Diego)
TransformerAuto EncoderTime SeriesBiomedical Data
🎯 What it does: Propose Coupled Transformer Autoencoder (CTAE), which learns nonlinear, non-stationary, long-range temporal latent variables in multi-region neural recordings through Transformer encoder/decoder, while orthogonally separating shared and private dynamics within the same framework.
Coupling Experts and Routers in Mixture-of-Experts via an Auxiliary Loss
Ang Lv (Renmin University of China), Siyuan Qiao (Bytedance Seed)
Mixture of ExpertsText
🎯 What it does: Proposed an expert-router coupling loss (ERC loss), aiming to enhance the coupling between the router's decisions and expert capabilities in Mixture-of-Experts (MoE) models through a lightweight auxiliary loss.
Covariate-Guided Clusterwise Linear Regression for Generalization to Unseen Data
Dohyun Bu (Korea Advanced Institute of Science and Technology), Jong-Seok Lee (Korea Advanced Institute of Science and Technology)
OptimizationExplainability and InterpretabilityMixture of ExpertsTabularBenchmark
🎯 What it does: Propose an end-to-end covariance-guided clustering linear regression framework (CG-CLR), which simultaneously learns sample allocation rules and multiple local linear regressors to achieve single-point prediction for unseen data
CP-Agent: Context‑Aware Multimodal Reasoning for Cellular Morphological Profiling under Chemical Perturbations
Yuxin Zhang (University of Hong Kong), Kevin Kin-Man Tsia (University of Hong Kong)
Explainability and InterpretabilityDrug DiscoveryLarge Language ModelAgentic AIVision Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Proposed CP-Agent, an intelligent agent combining context-aware CLIP with a multimodal large language model to generate explainable reports on drug-induced cell morphology changes.
CPiRi: Channel Permutation-Invariant Relational Interaction for Multivariate Time Series Forecasting
Jiyuan Xu (Zhejiang University of Finance and Economics), Shuai Zhang (Zhejiang University of Finance and Economics)
TransformerTime Series
🎯 What it does: Propose a multivariate time series forecasting framework CPiRi based on channel permutation invariance (CPI), addressing the dependency of traditional models on channel order and the neglect of interactions in independent models.
CPQS-Tuning: A Model Self-Perception-Based Data Filtering Algorithm for Efficient Instruction Fine-Tuning
Yi Ren (Nanjing University), Diandong Liu (Shaanxi University of Science & Technology)
Computational EfficiencyData-Centric LearningConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: This paper proposes a self-perception data filtering method called CPQS-Tuning based on the hidden states of large language models (LLMs) to enhance the efficiency and effectiveness of instruction fine-tuning.
CR-Net: Scaling Parameter-Efficient Training with Cross-Layer Low-Rank Structure
Boao Kong (Peking University), Kun Yuan (Peking University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose CR-Net, a parameter-efficient pre-training framework for large language models (LLMs) that leverages cross-layer low-rank activation residuals.
CreatiDesign: A Unified Multi-Conditional Diffusion Transformer for Creative Graphic Design
Hui Zhang (Fudan University), Yu-Gang Jiang (Fudan University)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelAuto EncoderMultimodality
🎯 What it does: Propose CreatiDesign, a unified multi-condition driven diffusion transformer for automated graphic design.
Credit-Budgeted ICPC-Style Coding: When Agents Must Pay for Every Decision
Lingfeng Zhou (Shanghai Jiao Tong University), Dequan Wang (Shanghai Jiao Tong University)
AI Code AssistantLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Designed and implemented USACOArena, an interactive coding arena based on ACM-ICPC, integrating programming tasks with a unified credit economy, requiring agents to make decisions under a limited budget.
CREPE: Controlling diffusion with REPlica Exchange
Jiajun He (University of Cambridge), Francisco Vargas (Xaira Therapeutics)
GenerationData SynthesisDiffusion modelMultimodality
🎯 What it does: Proposes a method called CREPE for inference-time control based on Replica Exchange (Parallel Tempering), generating samples that satisfy new constraints using pre-trained diffusion models without retraining.
Critic–Adviser–Reviser Cyclic Refinement: Towards High-Quality EMR Corpus Generation with LLMs
Chen Ning (Tsinghua University), Ji Wu (Tsinghua University)
Data SynthesisLarge Language ModelTextElectronic Health Records
🎯 What it does: Proposed a multi-stage iterative refinement framework based on LLM, named LLM-CARe, for generating synthetic electronic medical records that meet clinical quality standards
Critical attention scaling in long-context transformers
Shi Chen (MIT), Philippe Rigollet (MIT)
Computational EfficiencyRepresentation LearningTransformer
🎯 What it does: Theoretical analysis of the attention mechanism in long-context Transformers, proving that using polylogarithmic scaling (βₙ≈logⁿ) for attention scores can avoid rank-collapse.
Critical Confabulation: Can LLMs Hallucinate for Social Good?
Peiqi Sui (McGill University), Richard Jean So (Duke University)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the 'critical confabulation' framework, utilizing LLM-generated spontaneous hypothetical narratives as tools to fill gaps in historical archives, formulated as an open-ended narrative cloze task;
Critique-Coder: Enhancing Coder Models by Critique Reinforcement Learning
Chi Ruan (University Of Waterloo), Wenhu Chen (Vector Institute)
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed Critique Reinforcement Learning (CRL) for code generation tasks, integrated CRL training with RL to construct the CRITIQUE-CODER model.
Critique-RL: Training Language Models For Critiquing Through Two-Stage Reinforcement Learning
Zhiheng Xi (Fudan University), Xuanjing Huang (Fudan University)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed and implemented a two-stage reinforcement learning framework called Critique-RL for training critical language models that can evaluate and provide constructive feedback.
CroCoDiLight: Repurposing Cross-View Completion Encoders for Relighting
Alistair J Foggin (University of York), William A P Smith (University of York)
Image TranslationRestorationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: Reconstruct the latent space of CroCo to disentangle illumination information from scene intrinsic attributes, and utilize this disentanglement for illumination interpolation, shadow removal, and physical image decomposition.
CRONOS: Continuous time reconstruction for 4D medical longitudinal series
Nico Disch, Klaus Maier-Hein (German Cancer Research Center)
RestorationConvolutional Neural NetworkFlow-based ModelBiomedical DataOrdinary Differential Equation
🎯 What it does: Propose a unified continuous-time multi-view prediction framework named CRONOS, capable of predicting target volumes at any arbitrary time from multiple historical scans in 3D medical sequences.
Cross-ControlNet: Training-Free Fusion of Multiple Conditions for Text-to-Image Generation
Xiang Liu (Harbin Institute of Technology), Xianming Liu (Harbin Institute of Technology)
GenerationTransformerDiffusion modelMultimodality
🎯 What it does: Propose a training-free multimodal text-to-image generation framework that can achieve fine-grained control under conflicting and complementary conditions.
Cross-Domain Lossy Compression via Rate- and Classification-Constrained Optimal Transport
Nam Nguyen (Oregon State University), Bella Bose (Oregon State University)
ClassificationRestorationCompressionDomain AdaptationAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Study cross-domain lossless compression by leveraging a restricted optimal transport framework, mapping degraded inputs to target distributions while satisfying constraints on compression rate, classification error, and perceptual distance; derive closed-form DRC/RDC/DRPC expressions for Bernoulli and Gaussian models under one-shot and asymptotic regimes, and implement and validate them in deep end-to-end compression models.
Cross-Domain Policy Optimization via Bellman Consistency and Hybrid Critics
Ming-Hong Chen (National Yang Ming Chiao Tung University), Ping-Chun Hsieh (National Yang Ming Chiao Tung University)
Domain AdaptationReinforcement LearningFlow-based Model
🎯 What it does: This paper proposes a cross-domain reinforcement learning framework called Q Avatar, which achieves reliable knowledge transfer by combining the Critic from the source domain and the target domain.
Cross-Embodiment Offline Reinforcement Learning for Heterogeneous Robot Datasets
Haruki Abe (University of Tokyo), Tatsuya Harada (University of Tokyo)
Robotic IntelligenceReinforcement LearningGraph
🎯 What it does: Combine offline reinforcement learning with cross-embodiment learning to pretrain a general policy on a multi-robot dataset
Cross-Modal Redundancy and the Geometry of Vision–Language Embeddings
Grégoire DHIMOÏLA (Brown University), Agustin Martin Picard (DEEL IRT Saint Exupéry)
Representation LearningVision Language ModelAuto EncoderMultimodality
🎯 What it does: Study the geometric structure of shared embedding spaces in vision-language models, propose the Iso-Energy hypothesis, and extract cross-modal concepts through aligned sparse autoencoders.
Cross-Timestep: 3D Diffusion Model with Trans-temporal Memory LSTM and Adaptive Priori Decoding Strategy for Medical Segmentation
Shangqian Wu (Central South University), Lei Deng (Central South University)
SegmentationRecurrent Neural NetworkDiffusion modelBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposed the Cross-Timestep framework to address the initial phase collapse problem in 3D medical image segmentation
Cross-Tokenizer Likelihood Scoring Algorithms for Language Model Distillation
Buu Phan (University of Toronto), Karen Ullrich (Meta AI)
Knowledge DistillationTransformerTextBenchmark
🎯 What it does: This paper proposes a cross-tokenizer likelihood scoring algorithm, enabling language models to accurately evaluate sequence probabilities and perform knowledge distillation when using different BPE vocabularies.
CrossPL: Systematic Evaluation of Large Language Models for Cross Programming Language Interoperating Code Generation
zhanhang xiong, Wenhai Wang (Zhejiang University)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Created the CrossPL benchmark to evaluate LLMs' ability in cross-language interoperability (IPC and FFI) code generation, and systematically assessed the performance of 20 LLMs on this benchmark.
CryoLVM: Self-supervised Learning from Cryo-EM Density Maps with Large Vision Models
Weining Fu (Tsinghua University), Qiangfeng Cliff Zhang (Tsinghua University)
RestorationSuper ResolutionRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningContrastive LearningBiomedical Data
🎯 What it does: Propose CryoLVM, which leverages the JEPA self-supervised framework and SCUNet backbone to learn structural representations of 3D cryo-EM density maps, and fine-tunes on three downstream tasks: density map sharpening, super-resolution, and missing wedge recovery.
CryoNet.Refine: A One-step Diffusion Model for Rapid Refinement of Structural Models with Cryo-EM Density Map Restraints
Fuyao Huang (State Key Laboratory of Membrane Biology-Membrane Structure and Artificial Intelligence Biology Branch), Qiangfeng Cliff Zhang (Tsinghua University)
OptimizationProtein Structure PredictionTransformerDiffusion modelBiomedical Data
🎯 What it does: Developed a first-order diffusion model, CryoNet.Refine, to rapidly approximate experimental cryo-EM density maps from initial atomic models and improve geometric structures.
CryoSplat: Gaussian Splatting for Cryo-EM Homogeneous Reconstruction
Suyi Chen (Stony Brook University), Haibin Ling (Westlake University)
Protein Structure PredictionGaussian SplattingBiomedical Data
🎯 What it does: Propose a self-consistent cryo-EM homogeneous reconstruction method called cryoSplat, which enables physically accurate three-dimensional potential reconstruction from raw particle images through direct initialization.
CSRv2: Unlocking Ultra-Sparse Embeddings
Lixuan Guo (Stony Brook University), Chenyu You (Stony Brook University)
Computational EfficiencyRepresentation LearningSupervised Fine-TuningAuto EncoderContrastive LearningImageTextBiomedical Data
🎯 What it does: Propose CSRv2, a training framework for extremely sparse embeddings (k≤4), achieving efficient and high-quality sparse embeddings through progressive k-annealing, supervised sparse contrastive learning, and optional full model fine-tuning.
CTBench: Cryptocurrency Time Series Generation Benchmark
Yihao Ang (National University of Singapore), Zhiyong Huang (National University of Singapore)
GenerationData SynthesisDiffusion modelFlow-based ModelAuto EncoderGenerative Adversarial NetworkTime SeriesBenchmarkFinance Related
🎯 What it does: Designed CTBench, a benchmark framework for cryptocurrency time series generation, providing data, dual-task evaluation, and multidimensional financial metrics.
CTC-DRO: Robust Optimization for Reducing Language Disparities in Speech Recognition
Martijn Bartelds (Stanford University), Karen Livescu (Toyota Technological Institute at Chicago)
RecognitionOptimizationTransformerAudio
🎯 What it does: Propose a group distribution robust optimization algorithm called CTC-DRO for multi-voice recognition, addressing the subgroup imbalance issue caused by the incomparability of CTC loss in traditional Group DRO.
Ctrl-World: A Controllable Generative World Model for Robot Manipulation
Yanjiang Guo (Stanford University), Chelsea Finn (Tsinghua University)
GenerationRobotic IntelligenceDiffusion modelWorld ModelVideo
🎯 What it does: Constructed a controllable multi-perspective generative world model, Ctrl-World, for evaluating and enhancing general robot strategies in an imagined space.
CTRL&SHIFT: High-quality Geometry-Aware Object Manipulation in Visual Generation
Penghui Ruan (Hong Kong Polytechnic University), Yuhui Shi (Southern University of Science and Technology)
Image TranslationGenerationTransformerDiffusion modelImageVideo
🎯 What it does: Proposes the Ctrl&Shift framework based on diffusion models, enabling geometrically consistent object movement and perspective control in images and videos without requiring explicit 3D reconstruction.
CubeBench: Diagnosing Interactive, Long-Horizon Physical Intelligence under Partial Observations
Huan-ang Gao (Tsinghua University), Mengdi Wang (Princeton University)
Large Language ModelAgentic AISequentialBenchmark
🎯 What it does: This paper proposes CubeBench, a generative benchmark centered on the Rubik's cube, designed to diagnose the capabilities of LLM agents in spatial reasoning, long-sequence state tracking, and partially observed exploration.
CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning
Xiaoya Li (DeepReinforce Team), Chris Shum (DeepReinforce Team)
OptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: This paper proposes CUDA-L1, an automated CUDA optimization framework based on contrastive reinforcement learning, which can generate accelerated CUDA code without relying on manual tuning.
Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment Dataset
Lily H Zhang (Meta), Maximilian Nickel (Meta)
Data-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Explores the diversity gap of LLMs in multicultural and multivalued preference settings, and proposes negative correlation sampling (NC sampling) to generate diverse candidate answers, aiming to collect more representative multilingual preference data.
Culture In a Frame: C$^3$B as a Comic-Based Benchmark for Multimodal Culturally Awareness
Yuchen Song (Harbin Institute of Technology), Tiejun Zhao (Harbin Institute of Technology)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes C³B — a multimodal, multilingual, and multicultural benchmark for evaluating cultural awareness based on comics;
Culture in Action: Evaluating Text-to-Image Models through Social Activities
Sina Malakouti (University of Pittsburgh), Adriana Kovashka (University of Pittsburgh)
GenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Investigated the authenticity and accuracy evaluation of text-to-image models across different cultural social activities.
CUPID: A Plug-in Framework for Joint Aleatoric and Epistemic Uncertainty Estimation with a Single Model
Xinran Xu (Nanyang Technological University), Xiuyi Fan (Nanyang Technological University)
ClassificationSuper ResolutionAnomaly DetectionGenerative Adversarial NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a pluggable plugin called CUPID, which can jointly estimate the aleatoric and epistemic uncertainties of deep learning models without modifying or retraining the base model.
Curation Leaks: Membership Inference Attacks against Data Curation for Machine Learning
Dariush Wahdany (CISPA Helmholtz Center for Information Security), Franziska Boenisch (CISPA Helmholtz Center for Information Security)
Safty and PrivacyImageMultimodality
🎯 What it does: This paper systematically evaluates and attacks privacy leakage risks in the data curation process, designing membership inference attacks for three stages: curation scores, curated subsets, and the final model, and verifies the effectiveness of differential privacy (DP) defenses.
CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs
Yongcheng Zeng (Institute of Automation, Chinese Academy of Sciences), Jun Wang (University College London)
Computational EfficiencyLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose an adaptive curriculum learning framework called CurES based on gradient analysis and Bayesian posterior inference to enhance the training efficiency of reasoning large language models (LLMs).
Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning
Shubham Parashar (Texas A&M University), Shuiwang Ji (Texas A&M University)
TransformerReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes E2H Reasoner, a post-training method for large language models (LLMs) based on curriculum reinforcement learning (CRL), which enhances the reasoning ability of language models through task scheduling from easy to difficult.
Curse of Slicing: Why Sliced Mutual Information is a Deceptive Measure of Statistical Dependence
Alexander Semenenko (Applied AI Institute), Alexey Frolov (Applied AI Institute)
Contrastive LearningImageTabular
🎯 What it does: Systematically evaluate slice mutual information (SMI), revealing its defects in high-dimensional data such as saturation, susceptibility to noise, preference for redundant information, and failure to satisfy the data processing inequality;
Curvature-Guided Task Synergy for Skeleton based Temporal Action Segmentation
Guozhang Li (Beijing Normal University), Hua Huang (Beijing Normal University)
SegmentationGraph Neural NetworkTransformerMixture of ExpertsGraph
🎯 What it does: Propose a skeleton-based temporal action segmentation framework named CurvSeg, which utilizes curvature-guided task collaboration and dual-expert Mixture of Experts (MoE) to achieve interactive optimization between classification and boundary localization.
Customizing Visual Emotion Evaluation for MLLMs: An Open-vocabulary, Multifaceted, and Scalable Approach
Daiqing Wu (Chinese Academy of Sciences), Yu ZHOU
Large Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Propose the Emotion Statement Judgment task and its automated annotation pipeline INSETS, constructing and publicly releasing the MVEI benchmark to evaluate the ability of multimodal large language models (MLLM) in visual emotion understanding.
Cut Less, Fold More: Model Compression through the Lens of Projection Geometry
Olga Saukh (Graz University of Technology), Lothar Thiele (Swiss Data Science Center)
CompressionComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: Propose viewing structured pruning and model folding as orthogonal projections in the parameter space, investigate the geometric relationship between the two compression methods, and provide theoretical proofs.
Cutting the Skip: Training Residual-Free Transformers
Yiping Ji (Adelaide University), Simon Lucey (Adelaide University)
ClassificationSegmentationRetrievalTransformerImage
🎯 What it does: Propose a theory-driven initialization scheme that enables Transformers to stably train without residual (skip) connections
Cyber-Zero: Training Cybersecurity Agents without Runtime
Terry Yue Zhuo (Monash University), Zijian Wang (Meta Superintelligence Labs)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the CYBER-ZERO framework, which generates high-quality proxy trajectories without a runtime environment by leveraging publicly available CTF writing to train LLM performance in cybersecurity tasks.
CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale
Zhun Wang (University Of California Berkeley), Dawn Song (University Of California Berkeley)
AI Code AssistantLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Developed the CyberGym benchmark, collecting 1,507 real-world vulnerability instances to evaluate the ability of AI agents to generate reproducible PoCs based solely on vulnerability descriptions and codebases.
CyclicReflex: Improving Reasoning Models via Cyclical Reflection Token Scheduling
Chongyu Fan (Michigan State University), Sijia Liu (Michigan State University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a training-free decoding strategy called CyclicReflex, which dynamically balances thinking and decision-making by periodically adjusting reflection tokens during the reasoning process.
CylinderSplat: 3D Gaussian Splatting with Cylindrical Triplanes for Panoramic Novel View Synthesis
Qiwei Wang, Yujiao Shi
GenerationNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Proposes CylinderSplat, a dual-branch feedforward panoramic 3D Gaussian profile framework capable of generating realistic novel view images from single-view or sparse-view inputs.
D-AR: Diffusion via Autoregressive Models
Ziteng Gao (National University of Singapore), Mike Zheng Shou (National University of Singapore)
GenerationTransformerDiffusion modelFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Propose the D-AR framework, which reformulates the pixel-level diffusion process as a standard next-token prediction task; design an ordered diffusion tokenizer to map images into 1D discrete token sequences, then use the Llama decoder to generate these tokens, instantly decoding them into diffusion steps during generation to achieve streaming pixel generation.
D-REX: Differentiable Real-to-Sim-to-Real Engine for Learning Dexterous Grasping
Haozhe Lou (University of Southern California), Yue Wang (University of Southern California)
Robotic IntelligenceVision Language ModelGaussian SplattingVideo
🎯 What it does: Built a differentiable Real-to-Sim-to-Real framework called D-REX for identifying object quality from robot videos and training force-aware grasping strategies.
D&R: Recovery-based AI-Generated Text Detection via a Single Black-box LLM Call
Yuxia Sun (Jinan University), Jingcai Guo (Hong Kong Polytechnic University)
Anomaly DetectionTransformerLarge Language ModelText
🎯 What it does: Propose a zero-shot AI text detection framework named Disrupt-and-Recover (D&R), which utilizes model-agnostic Within-Chunk Shuffling and completes recovery through a single black-box LLM call, followed by determining the text origin via semantic and structural similarity.
d$^2$Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching
Yuchu Jiang (Southeast University), Xu Yang (Southeast University)
Computational EfficiencyLarge Language ModelDiffusion modelText
🎯 What it does: Accelerate the inference of discrete diffusion large language models (dLLM) through the Dual Adaptive Cache framework;
D$^2$GS: Depth-and-Density Guided Gaussian Splatting for Stable and Accurate Sparse-View Reconstruction
Meixi Song (Tsinghua University), Lu Qi (Insta360 Research)
Depth EstimationNeural Radiance FieldGaussian SplattingImage
🎯 What it does: This paper improves 3D Gaussian Splatting under sparse views by proposing two modules: Depth-and-Density Guided Dropout and Distance-Aware Fidelity Enhancement, enhancing the stability and quality of sparse view reconstruction.
D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI
Suhwan Choi (MAUM.AI), Yunsung Lee (MAUM.AI)
Robotic IntelligenceTransformerVision-Language-Action ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Proposes an end-to-end framework D2E for transferring pre-training from desktop data to robot embedded AI, including data collection tool OWA Toolkit, a generalist inverse dynamics model Generalist-IDM, and a visual-action pre-training method VAPT, demonstrating its effectiveness in robotic manipulation and navigation tasks.
DA$^{2}$: Depth Anything in Any Direction
Haodong Li (Tencent Hunyuan), Chunchao Guo (Tencent Hunyuan)
Depth EstimationTransformerImage
🎯 What it does: Propose the DA² framework to achieve end-to-end, unsupervised, zero-shot depth (scale-invariant distance) estimation for full panoramic (360°×180°) views, balancing high accuracy and strong generalization;
DAComp: Benchmarking Data Agents across the Full Data Intelligence Lifecycle
Fangyu Lei (Chinese Academy of Sciences), Kang Liu (Chinese Academy of Sciences)
TransformerLarge Language ModelAgentic AITextTabularBenchmark
🎯 What it does: Proposed and implemented the DAComp benchmark, covering two tasks: warehouse-level data engineering (DE) and open data analysis (DA), aiming to evaluate the end-to-end capabilities of LLM agents throughout the data intelligence lifecycle.
DADA: Dual Averaging with Distance Adaptation
Mohammad Moshtaghifar (University of British Columbia), Sebastian U Stich
Optimization
🎯 What it does: Proposed Dual Averaging with Distance Adaptation (DADA), a self-adaptive and general convex optimization gradient method.
DAG-Math: Graph-of-Thought Guided Mathematical Reasoning in LLMs
Yuanhe Zhang (University of Warwick), Fanghui Liu (Shanghai Jiao Tong University)
Graph Neural NetworkPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a mathematics reasoning framework DAG-MATH based on task-specific directed acyclic graphs (DAGs), and define the metrics of logical proximity and perfect reasoning rate.
DAK-UCB: Diversity-Aware Prompt Routing for LLMs and Generative Models
Donya Jafari (Sharif University of Technology), Farzan Farnia (Chinese University of Hong Kong)
GenerationLarge Language ModelPrompt EngineeringMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes an online multi-model selection framework called DAK-UCB, which can actively enhance output diversity in generation tasks while maintaining generation quality (e.g., CLIP-Score), and further introduces Mixture-DAK-UCB to achieve prompt-based hybrid model selection.
DAMR: Efficient and Adaptive Context-Aware Knowledge Graph Question Answering with LLM-Guided MCTS
Yingxu Wang (Mohamed bin Zayed University of Artificial Intelligence), Nan Yin (City University of Hong Kong)
Computational EfficiencyTransformerLarge Language ModelGraphChain-of-Thought
🎯 What it does: Proposes a framework named DAMR that performs dynamic and adaptive multi-hop reasoning in knowledge graph question answering using LLM-guided MCTS.